Xtlogit, fe : huge drop in variables

#1
Hello,

I'm estimating happiness levels of working women based on their career and marital aspects:

xtlogit happy jbsat_all jbsat_hrs finance chsee married##numchild if (fworking==1), fe

Where fworking is number of working females. Initially, my sample has about 53k observations:

fworking | Freq. Percent Cum.
------------+-----------------------------------
0 | 16,713 23.84 23.84
1 | 53,388 76.16 100.00
------------+-----------------------------------
Total | 70,101 100.00

But when running the xtlogit, fe --> I only seem to have 2000 observations left:

note: multiple positive outcomes within groups encountered.
note: 4089 groups (7694 obs) dropped because of all positive or
all negative outcomes.

Conditional fixed-effects logistic regression Number of obs = 2352
Group variable: pid Number of groups = 877

Obs per group: min = 2
avg = 2.7
max = 3

LR chi2(7) = 70.35
Log likelihood = -815.18271 Prob > chi2 = 0.0000

------------------------------------------------------------------------------
happy | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
jbsat_all | .866748 .1478742 5.86 0.000 .5769199 1.156576
jbsat_hrs | .200327 .1389541 1.44 0.149 -.072018 .4726719
finance | .7282601 .1909593 3.81 0.000 .3539868 1.102533
chsee | -.0181798 .012801 -1.42 0.156 -.0432693 .0069098
1.married | -.3143205 .2562828 -1.23 0.220 -.8166256 .1879846
1.numchild | -.641125 .2826098 -2.27 0.023 -1.19503 -.0872199
|
married#|
numchild |
1 1 | .5428987 .317085 1.71 0.087 -.0785765 1.164374
------------------------------------------------------------------------------

Could anyone perhaps let me know if I'm doing something wrong, or are there some variables that the fixed effects naturally drops such as the 4000 groups that were dropped at the beginning? I read somewhere that this is because there is no variability within the subject so they were dropped, so the women were either employed in all time periods or not employed in all time periods.

If so, is there any solution to this? I think 53,000 to 2,000 observations seems to be a huge error on my part.
 

bukharin

RoboStataRaptor
#2
Yes the 4089 groups dropped at the beginning is a consequence of the analysis method. However, this only explains the loss of 7694 observations, which means that you have still lost over 45,000 observations.

Probably the most likely explanation is missing data - observations will only be included in the model if they're not missing any data for any of the variables included in the model. As a first step you could try:
misstable summarize pid happy jbsat_all jbsat_hrs finance chsee married numchild